-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathomnisphero_mil.py
1486 lines (1303 loc) · 77.8 KB
/
omnisphero_mil.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
import math
import os
import random
import sys
import time
from datetime import datetime
from sys import getsizeof
import matplotlib.pyplot as plt
import numpy as np
from torch.utils.data import DataLoader
import multiprocessing
import socket
import hardware
import loader
import mil_metrics
import models
import omnisphero_mining
import predict_batch
import r
import torch_callbacks
import video_render_ffmpeg
from util import log
from util import paths
from util import sample_preview
from util import utils
from util.omnisphero_data_loader import OmniSpheroAugmentedDataLoader
from util.omnisphero_data_loader import OmniSpheroDataLoader
from util.paths import default_out_dir_unix_base
from util.paths import training_metrics_live_dir_name
from util.utils import line_print
from util.utils import shuffle_and_split_data
# setting env before importing torch
os.environ['CUDA_LAUNCH_BLOCKING'] = "1"
import torch
from torch.optim import Optimizer
# Just for pytorch: Setting the print-options so pytorch tensors are displayed in more detail
torch.set_printoptions(sci_mode=False, precision=8)
# On Windows, if there's not enough RAM:
# https://github.com/Spandan-Madan/Pytorch_fine_tuning_Tutorial/issues/10
# normalize_enum is an enum to determine normalisation as follows:
# 0 = no normalisation
# 1 = normalize every cell between 0 and 255 (8 bit)
# 2 = normalize every cell individually with every color channel independent
# 3 = normalize every cell individually with every color channel using the min / max of all three
# 4 = normalize every cell but with bounds determined by the brightest cell in the bag
# 5 = z-score every cell individually with every color channel independent
# 6 = z-score every cell individually with every color channel using the mean / std of all three
# 7 = z-score every cell individually with every color channel independent using all samples in the bag
# 8 = z-score every cell individually with every color channel using the mean / std of all three from all samples in the bag
normalize_enum_default = 3
max_workers_default = 5
def train_model(
# Basic training data params
training_label: str, source_dirs: [str], image_folder: str,
# Model fitting params
loss_function: str, device_ordinals: [int],
epochs: int = 3, max_workers: int = max_workers_default, normalize_enum: int = normalize_enum_default,
out_dir: str = None, gpu_enabled: bool = False,
shuffle_data_loaders: bool = True, model_enable_attention: bool = False, model_use_max: bool = True,
global_log_dir: str = None, optimizer: str = 'adam', include_line_fit_in_metrics_after_training: bool = True,
# Clamp Loss function
clamp_min: float = None, clamp_max: float = None,
# Callback configurations
stop_when_spiking_loss: bool = True,
early_stopping_enabled: bool = True,
halve_lr_enabled: bool = True, initial_lr_override: float = None,
# Tile shuffling
loading_preview_rate: float = 0.5,
repack_percentage: float = 0.0,
positive_bag_min_samples: int = None,
# Tile Constraints (How many Nuclei / Oligos / Neurons must be at least in a sample?)
tile_constraints_0: [int] = loader.default_tile_constraints_none,
tile_constraints_1: [int] = loader.default_tile_constraints_none,
# Well indices for labels. When a bag is loaded from a specific well index, the corresponding label is applied
label_0_well_indices=loader.default_well_indices_none,
label_1_well_indices=loader.default_well_indices_none,
force_balanced_batch: bool = False,
# Enable data augmentation?
augment_train: bool = False, augment_validation: bool = False,
write_whole_bag_previews: bool = False, write_sample_previews: bool = False,
# Training histogram bins override (if None, every histogram uses dynamic bin sizes)
hist_bins_override: int = None,
# Sigmoid Evaluation parameters
save_sigmoid_plot_interval: int = 5,
# Render sigmoid output as a video file after training?
sigmoid_video_render_enabled: bool = True, render_fps: int = video_render_ffmpeg.default_fps,
# What channels are enabled during loading?
channel_inclusions: [bool] = loader.default_channel_inclusions_all,
# Training / Validation Split percentages
data_split_percentage_validation: float = 0.35, data_split_percentage_test: float = 0.20,
# HNM Params
use_hard_negative_mining: bool = True, hnm_magnitude: float = 5.0, hnm_new_bag_percentage=0.25,
writing_metrics_enabled: bool = True,
# Test the model on the test data, after training?
testing_model_enabled: bool = True,
# Sigmoid validation dirs
sigmoid_validation_dirs: [str] = None, reserve_sigmoid_experiments_as_test_data: bool = True,
# reserve compounds for datasets
reserve_compound_train=None,
reserve_compound_test=None,
reserve_compound_validation=None,
# Limit the loading of neurospheres to specific quartiles for training data (Does not affect sigmoid data)
used_tile_quartiles=None,
# After training, should the model be run on input experiments again?
predict_training_data_afterwards: bool = False,
predict_sigmoid_data_afterwards: bool = False
):
if out_dir is None:
out_dir = source_dirs[0] + os.sep + 'training_results'
if not testing_model_enabled:
data_split_percentage_test = 0
if sigmoid_validation_dirs is None:
sigmoid_validation_dirs = []
reserve_sigmoid_experiments_as_test_data = False
sigmoid_evaluation_enabled = len(sigmoid_validation_dirs) > 0
repack_percentage = float(repack_percentage)
# mutable fix
if reserve_compound_train is None:
reserve_compound_train = []
if reserve_compound_test is None:
reserve_compound_test = []
if reserve_compound_validation is None:
reserve_compound_validation = []
if used_tile_quartiles is None:
used_tile_quartiles = loader.default_used_tile_quartiles.copy()
used_tile_quartiles_enum = utils.boolean_to_integer(used_tile_quartiles[0],
used_tile_quartiles[1],
used_tile_quartiles[2],
used_tile_quartiles[3])
# This param is unused and should not be "True"!
if type(label_0_well_indices) == list and type(label_1_well_indices) == list:
assert len(label_0_well_indices) > 0
assert len(label_1_well_indices) > 0
data_loader_cores = math.ceil(os.cpu_count() * 0.5 + 1)
data_loader_cores = int(min(data_loader_cores, 4))
# Setting up directories
out_dir = out_dir + os.sep + training_label + os.sep
loading_preview_dir = out_dir + os.sep + 'loading_previews' + os.sep
loading_preview_dir_whole_bag = loading_preview_dir + 'whole_bags' + os.sep
metrics_dir = out_dir + os.sep + 'metrics' + os.sep
sigmoid_validation_dir = out_dir + os.sep + training_metrics_live_dir_name + os.sep + 'sigmoid_live' + os.sep
os.makedirs(out_dir, exist_ok=True)
os.makedirs(metrics_dir, exist_ok=True)
os.makedirs(loading_preview_dir, exist_ok=True)
os.makedirs(loading_preview_dir_whole_bag, exist_ok=True)
os.makedirs(sigmoid_validation_dir, exist_ok=True)
if gpu_enabled:
log.write('Number of visible GPU devices: ' + str(torch.cuda.device_count()))
log.write('Model classification - Use Max: ' + str(model_use_max))
log.write('Model classification - Use Attention: ' + str(model_enable_attention))
log.write('Model classification - Use HNM: ' + str(use_hard_negative_mining))
log.write('R - Is pyRserve connection available: ' + str(r.has_connection(also_test_script=True)))
log.write('Saving logs and protocols to: ' + out_dir)
# Logging params and args
write_protocol(out_dir=out_dir, text='Start time: ' + utils.gct(), new_entry=True)
write_protocol(out_dir=out_dir, text='\nHostname: ' + str(socket.gethostname()))
write_protocol(out_dir=out_dir, text='\n\n == General Params ==')
write_protocol(out_dir=out_dir, text='\nSource dirs: ' + str(len(source_dirs)))
write_protocol(out_dir=out_dir, text='\nLoss function: ' + loss_function)
write_protocol(out_dir=out_dir, text='\nDevice ordinals: ' + str(device_ordinals))
write_protocol(out_dir=out_dir, text='\nR - Is pyRserve connection available: ' + str(r.has_connection()))
write_protocol(out_dir=out_dir, text='\nSaving sigmoid plot interval: ' + str(save_sigmoid_plot_interval))
write_protocol(out_dir=out_dir, text='\nEpochs: ' + str(epochs))
write_protocol(out_dir=out_dir, text='\nShuffle data loader: ' + str(shuffle_data_loaders))
write_protocol(out_dir=out_dir, text='\nMax File-Loader Workers: ' + str(max_workers))
write_protocol(out_dir=out_dir, text='\nMax "DataLoader" Workers: ' + str(data_loader_cores))
write_protocol(out_dir=out_dir, text='\nGPU Enabled: ' + str(gpu_enabled))
write_protocol(out_dir=out_dir, text='\nHNM Enabled: ' + str(use_hard_negative_mining))
write_protocol(out_dir=out_dir, text='\nClamp Min: ' + str(clamp_min))
write_protocol(out_dir=out_dir, text='\nClamp Max: ' + str(clamp_max))
write_protocol(out_dir=out_dir, text='\n\n == Loader Params ==')
write_protocol(out_dir=out_dir, text='\nNormalization Enum: ' + str(normalize_enum))
write_protocol(out_dir=out_dir,
text='\nNormalization Strategy: ' + loader.normalize_enum_descriptions[normalize_enum])
write_protocol(out_dir=out_dir, text='\nInvert Bag Labels: <deprecated>')
write_protocol(out_dir=out_dir, text='\nRepack: Percentage: ' + str(repack_percentage))
write_protocol(out_dir=out_dir, text='\nLoading Preview Rate: ' + str(loading_preview_rate))
write_protocol(out_dir=out_dir, text='\nRepack: Minimum Positive Samples: ' + str(positive_bag_min_samples))
write_protocol(out_dir=out_dir, text='\n\nWell indices label 0: ' + str(label_0_well_indices))
write_protocol(out_dir=out_dir, text='\nWell indices label 1: ' + str(label_1_well_indices))
write_protocol(out_dir=out_dir, text='\nForce Balanced Batch: ' + str(force_balanced_batch))
write_protocol(out_dir=out_dir, text='\nTile constraints explained: Minimum number of x [Nuclei, Oligos, Neurons]')
write_protocol(out_dir=out_dir, text='\nTile Constraints label 0: ' + str(tile_constraints_0))
write_protocol(out_dir=out_dir, text='\nTile Constraints label 1: ' + str(tile_constraints_1))
write_protocol(out_dir=out_dir, text='\nChannel Inclusions: ' + str(channel_inclusions))
write_protocol(out_dir=out_dir, text='\nUsed Tile Quartiles: ' + str(used_tile_quartiles))
write_protocol(out_dir=out_dir, text='\nUsed Tile Quartiles: ' + str(used_tile_quartiles_enum) + ' (Enum)')
write_protocol(out_dir=out_dir, text='\n\n == Data splitting Params: ==')
write_protocol(out_dir=out_dir,
text='\nData Split percentage: Validation: ' + str(data_split_percentage_validation))
write_protocol(out_dir=out_dir, text='\nData Split percentage: Test: ' + str(data_split_percentage_test))
write_protocol(out_dir=out_dir, text='\nNumber of Sigmoid Validation Dirs: ' + str(len(sigmoid_validation_dirs)))
write_protocol(out_dir=out_dir, text='\nSigmoid validation enabled: ' + str(sigmoid_evaluation_enabled))
write_protocol(out_dir=out_dir, text='\nSigmoid validation experiments reserved for test data: ' + str(
reserve_sigmoid_experiments_as_test_data))
write_protocol(out_dir=out_dir, text='\nCompounds to be put in training data: ' + str(reserve_compound_train))
write_protocol(out_dir=out_dir,
text='\nCompounds to be put in validation data: ' + str(reserve_compound_validation))
write_protocol(out_dir=out_dir, text='\nCompounds to be put in test data: ' + str(reserve_compound_test))
global_log_filename = None
local_log_filename = out_dir + os.sep + 'log.txt'
log.add_file(local_log_filename)
if global_log_dir is not None:
global_log_filename = global_log_dir + os.sep + 'log-' + training_label + '.txt'
os.makedirs(global_log_dir, exist_ok=True)
log.add_file(global_log_filename)
log.diagnose()
# PREPARING DATA AND DIRECTORIES
write_protocol(out_dir=out_dir, text='\n\n == Directories ==')
write_protocol(out_dir=out_dir, text='\nGlobal Log dir: ' + str(global_log_dir))
write_protocol(out_dir=out_dir, text='\nLocal Log dir: ' + str(local_log_filename))
write_protocol(out_dir=out_dir, text='\nOut dir: ' + str(out_dir))
write_protocol(out_dir=out_dir, text='\nMetrics dir: ' + str(metrics_dir))
write_protocol(out_dir=out_dir, text='\nPreview tiles: ' + str(loading_preview_dir))
write_protocol(out_dir=out_dir,
text='\nPredict training input data afterwards: ' + str(predict_training_data_afterwards))
write_protocol(out_dir=out_dir,
text='\nPredict sigmoid input data afterwards: ' + str(predict_sigmoid_data_afterwards))
print('==== List of Source Dirs: =====')
[print(str(p)) for p in source_dirs]
write_protocol(out_dir=out_dir, text='\n\n == GPU Status ==\n')
for line in hardware.print_gpu_status(silent=True):
line = str(line).strip()
write_protocol(out_dir=out_dir, text=line)
log.write(line + '\n')
########################
# LOADING SIGMOID DATA
########################
f = open(out_dir + 'sigmoid-validation.txt', 'w')
data_loader_sigmoid: DataLoader = None
X_metadata_sigmoid: [np.ndarray] = None
sigmoid_experiment_names = []
X_sigmoid = []
if sigmoid_evaluation_enabled:
f.write('Sigmoid validation dirs:' + str(sigmoid_validation_dirs))
log.write('##### LOADING SIGMOID VALIDATION DIRS ####')
X_sigmoid, _, _, _, X_metadata_sigmoid, _, _, sigmoid_experiment_names, _, errors_sigmoid, loaded_files_list_sigmoid = loader.load_bags_json_batch(
batch_dirs=sigmoid_validation_dirs,
max_workers=max_workers,
include_raw=True,
force_balanced_batch=False,
channel_inclusions=channel_inclusions,
constraints_0=tile_constraints_0,
constraints_1=tile_constraints_1,
used_tile_quartiles=loader.default_used_tile_quartiles,
label_0_well_indices=loader.default_well_indices_all,
label_1_well_indices=loader.default_well_indices_all,
normalize_enum=normalize_enum)
X_sigmoid = [np.einsum('bhwc->bchw', bag) for bag in X_sigmoid]
sigmoid_temp_entry = None
f.write('\n\nList of loaded sigmoid files:')
for sigmoid_temp_entry in loaded_files_list_sigmoid:
f.write('\n' + str(sigmoid_temp_entry))
# log.write('Loaded sigmoid file: ' + str(sigmoid_temp_entry))
f.write('\n\nList of sigmoid loading errors:')
for sigmoid_temp_entry in errors_sigmoid:
f.write('\n' + str(sigmoid_temp_entry))
log.write('Loading error: ' + str(sigmoid_temp_entry))
if not reserve_sigmoid_experiments_as_test_data:
sigmoid_experiment_names = []
del errors_sigmoid, loaded_files_list_sigmoid, sigmoid_temp_entry
else:
reserve_sigmoid_experiments_as_test_data = False
sigmoid_experiment_names = []
f.write('Not sigmoid validating.')
log.write('Sigmoid validating: Disabled.')
f.close()
##############################
# LOADING TRAINING DATA START
##############################
unrestricted_experiments_override = None
if sigmoid_evaluation_enabled and reserve_sigmoid_experiments_as_test_data and testing_model_enabled:
unrestricted_experiments_override = sigmoid_experiment_names
# TODO Write well / label mapping to protocol file!
log.write('##### LOADING TRAINING / VALIDATION DATA DIRS ####')
loading_start_time = datetime.now()
X, y, y_tiles, X_raw, X_metadata, bag_names, _, _, _, errors, loaded_files_list = loader.load_bags_json_batch(
batch_dirs=source_dirs,
max_workers=max_workers,
include_raw=True,
force_balanced_batch=force_balanced_batch,
channel_inclusions=channel_inclusions,
constraints_0=tile_constraints_0,
constraints_1=tile_constraints_1,
label_0_well_indices=label_0_well_indices,
label_1_well_indices=label_1_well_indices,
used_tile_quartiles=used_tile_quartiles,
unrestricted_experiments_override=unrestricted_experiments_override,
normalize_enum=normalize_enum)
X = [np.einsum('bhwc->bchw', bag) for bag in X]
X_raw = [np.einsum('bhwc->bchw', bag) for bag in X_raw]
# Hint: Dim should be (xxx, 3, 150, 150)
loading_time = utils.get_time_diff(loading_start_time)
log.write('Loading finished in: ' + str(loading_time))
# Finished loading. Printing errors and data
f = open(out_dir + 'loading-errors.txt', 'w')
for e in errors:
f.write(str(e))
f.write('\n')
f.close()
# Saving one random image from random bags to the disk
log.write('Writing loading preview samples to: ' + loading_preview_dir)
preview_indices_0 = np.where(np.asarray(y) == 0)[0]
np.random.shuffle(preview_indices_0)
preview_indices_0 = preview_indices_0[:-max(math.floor(len(preview_indices_0) * (loading_preview_rate * 0.55)), 1)]
preview_indices_1 = np.where(np.asarray(y) == 1)[0]
np.random.shuffle(preview_indices_1)
preview_indices_1 = preview_indices_1[:-max(math.floor(len(preview_indices_1) * (loading_preview_rate * 0.55)), 1)]
preview_indices = np.append(preview_indices_0, preview_indices_1)
preview_indices = list(preview_indices)
preview_indices.sort()
print('\n')
for i in range(len(preview_indices)):
# Not doing that on windows devices
if os.name == 'nt':
continue
# not doing that with disabled parameter
if not write_sample_previews:
continue
# TODO expose to outer loop
line_print('Writing loading preview: ' + str(i + 1) + '/' + str(len(preview_indices)), include_in_log=False)
current_x: np.ndarray = X[preview_indices[i]]
j = random.randint(0, current_x.shape[0] - 1)
current_bag_name = bag_names[preview_indices[i]]
preview_image_file_base = loading_preview_dir + 'preview_' + str(
preview_indices[i]) + '-' + current_bag_name + '-' + str(
j) + '_' + str(y[preview_indices[i]])
current_x = current_x[j]
current_x: np.ndarray = np.einsum('abc->cba', current_x)
if current_x.min() >= 0 and current_x.max() <= 1:
# Normalized Image
preview_image_file = preview_image_file_base + '.png'
sample_preview.save_normalized_rgb(current_x, preview_image_file)
if normalize_enum >= 5:
sample_preview.save_z_scored_image(current_x, dim_x=150, dim_y=150,
fig_titles=['r (Nuclei)', 'g (Oligos)', 'b (Neurites)'],
filename=preview_image_file_base + '-z.png',
vmin=-3.0, vmax=3.0, normalize_enum=normalize_enum)
del current_x
del preview_indices, preview_indices_0, preview_indices_1, unrestricted_experiments_override
print('\n')
# Calculating Bag Size and possibly inverting labels
X_s = str(utils.byteSizeString(utils.listToBytes(X)))
X_s_raw = str(utils.byteSizeString(utils.listToBytes(X_raw)))
y_s = str(utils.byteSizeString(getsizeof(y)))
log.write('Finished loading data. Number of bags: ' + str(len(X)) + '. Number of labels: ' + str(len(y)))
log.write("X-size in memory (after loading all data): " + str(X_s))
log.write("y-size in memory (after loading all data): " + str(y_s))
log.write("X-size (raw) in memory (after loading all data): " + str(X_s_raw))
write_protocol(out_dir=out_dir, text='\n\n == Loaded Data ==')
write_protocol(out_dir=out_dir, text='\nNumber of Bags: ' + str(len(X)))
write_protocol(out_dir=out_dir, text='\nBags with label 0: ' + str(len(np.where(np.asarray(y) == 0)[0])))
write_protocol(out_dir=out_dir, text='\nBags with label 1: ' + str(len(np.where(np.asarray(y) == 1)[0])))
write_protocol(out_dir=out_dir, text="\nX-size in memory: " + str(X_s))
write_protocol(out_dir=out_dir, text="\ny-size in memory: " + str(y_s))
# Printing more data
f = open(out_dir + 'loading-data-statistics.csv', 'w')
for i in range(len(loaded_files_list)):
f.write(str(i) + ';' + loaded_files_list[i] + '\n')
f.write('\n\nX-size in memory: ' + str(X_s))
f.write('\n\ny-size in memory: ' + str(y_s))
f.write('\n\nLoading time: ' + str(loading_time))
f.close()
del X_s, y_s, X_s_raw, f
if len(X) == 0:
log.write('WARNING: NO DATA LOADED')
write_protocol(out_dir=out_dir, text='\n\nWARNING: NO DATA LOADED')
return
# Data Augmentation
# TODO move this somewhere else?
f = open(out_dir + 'data-augmentation.txt', 'w')
f.write('## Augmentation Train:\n' + str(augment_train) + '\n\n')
f.write('## Augmentation Validation:\n' + str(augment_validation) + '\n\n')
f.close()
#######################
# SETTING UP DATASETS
#######################
X_sigmoid_overlap = None
y_sigmoid_overlap = None
y_tiles_sigmoid_overlap = None
bag_names_sigmoid_overlap = None
X_raw_sigmoid_overlap = None
X_train_overlap = None
y_train_overlap = None
y_tiles_train_overlap = None
bag_names_train_overlap = None
X_raw_train_overlap = None
X_test_overlap = None
y_test_overlap = None
y_tiles_test_overlap = None
bag_names_test_overlap = None
X_raw_test_overlap = None
X_validation_overlap = None
y_validation_overlap = None
y_tiles_validation_overlap = None
bag_names_validation_overlap = None
X_raw_validation_overlap = None
# Extracting experiment names for reserved compounds
reserve_experiments_validation = utils.get_experiment_names_for_compound(X_metadata=X_metadata,
compound_names=reserve_compound_validation)
reserve_experiments_train = utils.get_experiment_names_for_compound(X_metadata=X_metadata,
compound_names=reserve_compound_train)
reserve_experiments_test = utils.get_experiment_names_for_compound(X_metadata=X_metadata,
compound_names=reserve_compound_test)
f = open(out_dir + 'reserve-sigmoid-as-test-data.txt', 'w')
f.write('## Param sigmoid_evaluation_enabled: ' + str(sigmoid_evaluation_enabled) + '\n')
f.write(
'## Param reserve_sigmoid_experiments_as_test_data: ' + str(reserve_sigmoid_experiments_as_test_data) + '\n')
f.write('## Param testing_model_enabled: ' + str(testing_model_enabled) + '\n\n')
if sigmoid_evaluation_enabled and reserve_sigmoid_experiments_as_test_data and testing_model_enabled:
X, X_metadata, X_raw, y, y_tiles, bag_names, X_sigmoid_overlap, X_metadata_sigmoid_overlap, X_raw_sigmoid_overlap, y_sigmoid_overlap, y_tiles_sigmoid_overlap, bag_names_sigmoid_overlap = extract_experiment_names_from_loaded_data(
out_dir=out_dir, file_name_suffix='-sigmoid-experiments',
extraction_experiment_names=sigmoid_experiment_names,
X=X, X_raw=X_raw, y=y, y_tiles=y_tiles, bag_names=bag_names, X_metadata=X_metadata
)
else:
f.write('Not running.')
if len(reserve_experiments_validation) > 0:
X, X_metadata, X_raw, y, y_tiles, bag_names, X_validation_overlap, X_metadata_validation_overlap, X_raw_validation_overlap, y_validation_overlap, y_tiles_validation_overlap, bag_names_validation_overlap = extract_experiment_names_from_loaded_data(
out_dir=out_dir, file_name_suffix='-validation-experiments',
extraction_experiment_names=reserve_experiments_validation,
X=X, X_raw=X_raw, y=y, y_tiles=y_tiles, bag_names=bag_names, X_metadata=X_metadata
)
if len(reserve_experiments_test) > 0:
X, X_metadata, X_raw, y, y_tiles, bag_names, X_test_overlap, X_metadata_test_overlap, X_raw_test_overlap, y_test_overlap, y_tiles_test_overlap, bag_names_test_overlap = extract_experiment_names_from_loaded_data(
out_dir=out_dir, file_name_suffix='-test-experiments',
extraction_experiment_names=reserve_experiments_test,
X=X, X_raw=X_raw, y=y, y_tiles=y_tiles, bag_names=bag_names, X_metadata=X_metadata
)
if len(reserve_experiments_train) > 0:
X, X_metadata, X_raw, y, y_tiles, bag_names, X_train_overlap, X_metadata_train_overlap, X_raw_train_overlap, y_train_overlap, y_tiles_train_overlap, bag_names_train_overlap = extract_experiment_names_from_loaded_data(
out_dir=out_dir, file_name_suffix='-train-experiments',
extraction_experiment_names=reserve_experiments_train,
X=X, X_raw=X_raw, y=y, y_tiles=y_tiles, bag_names=bag_names, X_metadata=X_metadata
)
# After repacking, X_metadata is invalidated! Must be deleted now. For your own safety
del X_metadata
# Printing Bag Shapes
# Setting up bags for MIL
if repack_percentage > 0:
log.write('Repack percent: ' + str(
repack_percentage) +
'.That means, to build a positive bag, x% of positive samples will be added to a negative bag.')
print_bag_metadata(X, y, y_tiles, bag_names, file_name=out_dir + 'bags-pre-packed.csv')
X, X_raw, y, y_tiles, bag_names = loader.repack_bags_merge(X=X, X_raw=X_raw, y=y, bag_names=bag_names,
repack_percentage=repack_percentage,
positive_bag_min_samples=positive_bag_min_samples)
print_bag_metadata(X, y, y_tiles, bag_names, file_name=out_dir + 'bags-repacked.csv')
if sigmoid_evaluation_enabled and reserve_sigmoid_experiments_as_test_data:
print_bag_metadata(X_sigmoid_overlap, y_sigmoid_overlap, y_tiles_sigmoid_overlap, bag_names_sigmoid_overlap,
file_name=out_dir + 'bags-pre-packed_sigmoid_overlap.csv')
X_sigmoid_overlap, X_raw_sigmoid_overlap, y_sigmoid_overlap, y_tiles_sigmoid_overlap, bag_names_sigmoid_overlap = loader.repack_bags_merge(
X=X_sigmoid_overlap, X_raw=X_raw_sigmoid_overlap, y=y_sigmoid_overlap,
bag_names=bag_names_sigmoid_overlap,
repack_percentage=repack_percentage,
positive_bag_min_samples=positive_bag_min_samples)
print_bag_metadata(X_sigmoid_overlap, y_sigmoid_overlap, y_tiles_sigmoid_overlap, bag_names_sigmoid_overlap,
file_name=out_dir + 'bags-repacked_sigmoid_overlap.csv')
else:
print_bag_metadata(X, y, y_tiles, bag_names, file_name=out_dir + 'bags.csv')
if sigmoid_evaluation_enabled and reserve_sigmoid_experiments_as_test_data and testing_model_enabled:
print_bag_metadata(X_sigmoid_overlap, y_sigmoid_overlap, y_tiles_sigmoid_overlap, bag_names_sigmoid_overlap,
file_name=out_dir + 'bags-sigmoid-overlap.csv')
if len(reserve_experiments_validation) > 0:
print_bag_metadata(X_validation_overlap, y_validation_overlap, y_tiles_validation_overlap,
bag_names_validation_overlap,
file_name=out_dir + 'bags-validation-overlap.csv')
if len(reserve_experiments_train) > 0:
print_bag_metadata(X_train_overlap, y_train_overlap, y_tiles_train_overlap, bag_names_train_overlap,
file_name=out_dir + 'bags-train-overlap.csv')
if len(reserve_experiments_test) > 0:
print_bag_metadata(X_test_overlap, y_test_overlap, y_tiles_test_overlap, bag_names_test_overlap,
file_name=out_dir + 'bags-test-overlap.csv')
# Removing unnecessary sigmoid overlap data
del bag_names_sigmoid_overlap, X_raw_sigmoid_overlap
del bag_names_validation_overlap, X_raw_validation_overlap
del bag_names_test_overlap, X_raw_test_overlap
del bag_names_train_overlap, X_raw_train_overlap
#########################
# WRITING BAG PREVIEWS
#########################
# Writing whole bags to the disc
preview_indexes_positive = list(np.where(np.asarray(y) == 1)[0])
preview_indexes_negative = list(np.where(np.asarray(y) == 0)[0])
random.shuffle(preview_indexes_positive)
random.shuffle(preview_indexes_negative)
preview_indexes_negative = preview_indexes_negative[0:math.ceil(len(preview_indexes_negative) * 0.15)]
preview_indexes_positive = preview_indexes_positive[0:math.ceil(len(preview_indexes_negative) * 0.25)]
preview_indexes = preview_indexes_negative
preview_indexes.extend(preview_indexes_positive)
preview_indexes.sort()
log.write('Number of whole bag previews to save: ' + str(len(preview_indexes)) + '. -> ' + str(preview_indexes))
print('\n')
# Writing whole bag previews
if write_whole_bag_previews:
log.write('Starting to write whole bag previews.')
for i in range(len(X)):
preview_image_filename = loading_preview_dir_whole_bag + 'preview_' + str(i) + '-' + bag_names[
i] + '_' + str(
y[i]) + '_bag.png'
line_print(
'Writing whole bag loading preview: ' + str(i + 1) + '/' + str(
len(X)) + ' -> ' + preview_image_filename, include_in_log=False)
colored_tiles = []
image_width = None
image_height = None
if i in preview_indexes:
for rgb in X_raw[i]:
# Creating a deep copy so it's not overwritten
rgb = np.copy(rgb)
rgb = rgb.copy()
image_width, image_height = rgb[0].shape
rgb = np.einsum('abc->bca', rgb)
rgb = mil_metrics.outline_rgb_array(rgb, None, None, outline=2, override_colormap=[255, 255, 255])
colored_tiles.append(rgb)
if len(colored_tiles) > 0 and image_height is not None:
out_image = mil_metrics.fuse_image_tiles(images=colored_tiles, image_width=image_width,
image_height=image_height)
plt.imsave(preview_image_filename, out_image)
line_print('Saved: ' + preview_image_filename)
log.write('Finished writing previews.')
else:
log.write('But not saving whole bag previews, due to user choice.')
######################################
# Converted raw data into datasets
######################################
log.write('Converting bags into batches.')
dataset, input_dim = loader.convert_bag_to_batch(bags=X, labels=y, y_tiles=y_tiles)
dataset_sigmoid_overlap = None
dataset_train_overlap = None
dataset_test_overlap = None
dataset_validation_overlap = None
log.write('Finished converting bags into patches.')
log.write('Detected input dim: ' + str(input_dim))
if sigmoid_evaluation_enabled and reserve_sigmoid_experiments_as_test_data and testing_model_enabled:
dataset_sigmoid_overlap, _ = loader.convert_bag_to_batch(bags=X_sigmoid_overlap, labels=y_sigmoid_overlap,
y_tiles=y_tiles_sigmoid_overlap)
if len(reserve_experiments_validation) > 0:
dataset_validation_overlap, _ = loader.convert_bag_to_batch(bags=X_validation_overlap,
labels=y_validation_overlap,
y_tiles=y_tiles_validation_overlap)
if len(reserve_experiments_train) > 0:
dataset_train_overlap, _ = loader.convert_bag_to_batch(bags=X_train_overlap, labels=y_train_overlap,
y_tiles=y_tiles_train_overlap)
if len(reserve_experiments_test) > 0:
dataset_test_overlap, _ = loader.convert_bag_to_batch(bags=X_test_overlap, labels=y_test_overlap,
y_tiles=y_tiles_test_overlap)
del X, y, preview_indexes, preview_indexes_positive, preview_indexes_negative
del X_sigmoid_overlap, y_sigmoid_overlap, y_tiles_sigmoid_overlap
del X_test_overlap, y_test_overlap, y_tiles_test_overlap
del X_train_overlap, y_train_overlap, y_tiles_train_overlap
del X_validation_overlap, y_validation_overlap, y_tiles_validation_overlap
##########################
# RANDOM TRAIN TEST SPLIT
##########################
log.write('Shuffling and splitting data into train and val set')
test_data = []
training_data, validation_data = shuffle_and_split_data(dataset,
additional_dataset_training=dataset_train_overlap,
additional_dataset_validation=dataset_validation_overlap,
split_percentage=data_split_percentage_validation)
if data_split_percentage_test is not None and data_split_percentage_test > 0:
combined_test_dataset = []
if dataset_sigmoid_overlap is not None:
combined_test_dataset.extend(dataset_sigmoid_overlap.copy())
if dataset_test_overlap is not None:
combined_test_dataset.extend(dataset_test_overlap.copy())
training_data, test_data = shuffle_and_split_data(dataset=training_data,
additional_dataset_training=dataset_train_overlap,
additional_dataset_validation=combined_test_dataset,
split_percentage=data_split_percentage_test)
del combined_test_dataset
# Clearing up memory from all these datasets
del dataset
del dataset_sigmoid_overlap, dataset_train_overlap, dataset_validation_overlap
f = open(out_dir + 'data-distribution.txt', 'w')
training_data_tiles: int = sum([training_data[i][0].shape[0] for i in range(len(training_data))])
validation_data_tiles: int = sum([validation_data[i][0].shape[0] for i in range(len(validation_data))])
log.write('Training data: ' + str(training_data_tiles) + ' samples over ' + str(len(training_data)) + ' bags.')
log.write(
'Validation data: ' + str(validation_data_tiles) + ' samples over ' + str(len(validation_data)) + ' bags.')
write_protocol(out_dir=out_dir, text='\nTraining data: ' + str(training_data_tiles) + ' samples over ' + str(
len(training_data)) + ' bags.')
write_protocol(out_dir=out_dir, text='\nValidation data: ' + str(validation_data_tiles) + ' samples over ' + str(
len(validation_data)) + ' bags.')
f.write('Training data: ' + str(training_data_tiles) + ' samples over ' + str(len(training_data)) + ' bags.\n')
f.write(
'Validation data: ' + str(validation_data_tiles) + ' samples over ' + str(len(validation_data)) + ' bags.\n')
if data_split_percentage_test is not None:
test_data_tiles: int = sum([test_data[i][0].shape[0] for i in range(len(test_data))])
log.write('Test data: ' + str(test_data_tiles) + ' samples over ' + str(len(test_data)) + ' bags.')
f.write('Test data: ' + str(test_data_tiles) + ' samples over ' + str(len(test_data)) + ' bags.\n')
write_protocol(out_dir=out_dir,
text='\nTest data: ' + str(test_data_tiles) + ' samples over ' + str(len(test_data)) + ' bags.')
f.close()
# Loading Hardware Device
device = hardware.get_hardware_device(gpu_preferred=gpu_enabled)
log.write('Selected device: ' + str(device))
# Loader args
loader_kwargs = {}
data_loader_pin_memory = False
if torch.cuda.is_available():
# model.cuda()
loader_kwargs = {'num_workers': data_loader_cores, 'pin_memory': data_loader_pin_memory}
#############################
# SETTING UP DATA LOADERS
#############################
# Data Generators
test_dl = None
if augment_train:
train_dl = OmniSpheroAugmentedDataLoader(training_data, batch_size=1, shuffle=shuffle_data_loaders,
transform_enabled=augment_train,
transform_data_saver=False, **loader_kwargs)
else:
train_dl = DataLoader(training_data, batch_size=1, shuffle=shuffle_data_loaders, **loader_kwargs)
if augment_validation:
validation_dl = OmniSpheroAugmentedDataLoader(validation_data, batch_size=1,
transform_enabled=augment_validation,
transform_data_saver=False,
shuffle=shuffle_data_loaders, **loader_kwargs)
else:
validation_dl = DataLoader(validation_data, batch_size=1, shuffle=shuffle_data_loaders, **loader_kwargs)
if data_split_percentage_test is not None:
test_dl = DataLoader(test_data, batch_size=1, shuffle=shuffle_data_loaders, **loader_kwargs)
del validation_data, test_data
if sigmoid_evaluation_enabled:
dataset_sigmoid, _ = loader.convert_bag_to_batch(bags=X_sigmoid, labels=None, y_tiles=None)
data_loader_sigmoid = DataLoader(dataset_sigmoid, batch_size=1, shuffle=False, **loader_kwargs)
del dataset_sigmoid, X_sigmoid
################
# MODEL START
################
# Setting up Model
log.write('Setting up model.')
log.write('Device Ordinals: ' + str(device_ordinals))
accuracy_function = 'binary'
model = models.BaselineMIL(input_dim=input_dim, device=device,
use_max=model_use_max,
enable_attention=model_enable_attention,
device_ordinals=device_ordinals,
loss_function=loss_function,
accuracy_function=accuracy_function)
log.write('Created blank model.')
# Saving the raw version of this model
untrained_model_path = out_dir + os.sep + 'model.h5'
log.write('Saving soon to be trained model to: ' + untrained_model_path)
torch.save(model.state_dict(), out_dir + 'model.pt')
torch.save(model, untrained_model_path)
log.write('Saved.')
log.write('Setting up optimizer.')
model_optimizer, initial_lr = models.choose_optimizer(model, selection=optimizer)
if initial_lr_override is not None:
initial_lr = initial_lr_override
initial_lr = float(initial_lr)
log.write('Finished loading data and model')
log.write('Optimizer: ' + str(model_optimizer))
log.write('Initial learning rate: ' + str(initial_lr))
# Callbacks
callbacks = []
hnm_callbacks = []
early_stopping_epoch_threshold = int(epochs / 5 + 1)
halve_lr_epoch_threshold = int(early_stopping_epoch_threshold / 2 + 1)
if stop_when_spiking_loss:
callbacks.append(torch_callbacks.SpikingLossCallback(loss_max=40.0))
hnm_callbacks.append(torch_callbacks.SpikingLossCallback(loss_max=40.0))
if early_stopping_enabled:
hnm_callbacks.append(torch_callbacks.EarlyStopping(epoch_threshold=early_stopping_epoch_threshold))
callbacks.append(torch_callbacks.EarlyStopping(epoch_threshold=early_stopping_epoch_threshold))
if halve_lr_enabled:
hnm_callbacks.append(
torch_callbacks.ReduceLearnRate(epoch_threshold=halve_lr_epoch_threshold, initial_lr=initial_lr,
optimizer=model_optimizer))
callbacks.append(
torch_callbacks.ReduceLearnRate(epoch_threshold=halve_lr_epoch_threshold, initial_lr=initial_lr,
optimizer=model_optimizer))
write_protocol(out_dir=out_dir, text='\n\n == Model Information==')
write_protocol(out_dir=out_dir, text='\nDevice Ordinals: ' + str(device_ordinals))
write_protocol(out_dir=out_dir, text='\nInitial LR: ' + str(initial_lr))
write_protocol(out_dir=out_dir, text='\nInput dim: ' + str(input_dim))
write_protocol(out_dir=out_dir, text='\ntorch Device: ' + str(device))
write_protocol(out_dir=out_dir, text='\nLoss Function: ' + str(loss_function))
write_protocol(out_dir=out_dir, text='\nAccuracy Function: ' + str(accuracy_function))
write_protocol(out_dir=out_dir, text='\nModel classification - Use Max: ' + str(model_use_max))
write_protocol(out_dir=out_dir, text='\nModel classification - Use Attention: ' + str(model_enable_attention))
write_protocol(out_dir=out_dir, text='\n\nData Loader - Cores: ' + str(data_loader_cores))
write_protocol(out_dir=out_dir, text='\nData Loader - Pin Memory: ' + str(data_loader_pin_memory))
write_protocol(out_dir=out_dir, text='\n\nCallback Count: ' + str(len(callbacks)))
write_protocol(out_dir=out_dir, text='\nCallbacks: ' + str(callbacks))
for c in callbacks:
write_protocol(out_dir=out_dir, text='\n' + c.describe())
write_protocol(out_dir=out_dir,
text='\n\nEarly stopping threshold (enabled: ' + str(early_stopping_enabled) + '): ' + str(
early_stopping_epoch_threshold))
write_protocol(out_dir=out_dir,
text='\nLearn Rate reduction threshold (enabled: ' + str(halve_lr_enabled) + '): ' + str(
halve_lr_epoch_threshold))
write_protocol(out_dir=out_dir, text='\n\nBuilt Optimizer: ' + str(model_optimizer))
# Printing the model
f = open(out_dir + os.sep + 'model.txt', 'w')
f.write(str(model))
f.close()
################
# TRAINING START
################
log.write(
'Start of training for ' + str(epochs) + ' epochs. Devices: ' + str(device_ordinals) + '. GPU enabled: ' + str(
gpu_enabled))
log.write('Training: "' + training_label + '"!')
history, history_keys, model_save_path_best = models.fit(model=model,
optimizer=model_optimizer,
epochs=epochs,
training_data=train_dl,
validation_data=validation_dl,
out_dir_base=out_dir,
bag_names=bag_names,
checkpoint_interval=None,
sigmoid_video_render_enabled=sigmoid_video_render_enabled,
render_fps=render_fps,
hist_bins_override=hist_bins_override,
sigmoid_evaluation_enabled=sigmoid_evaluation_enabled,
save_sigmoid_plot_interval=save_sigmoid_plot_interval,
data_loader_sigmoid=data_loader_sigmoid,
X_metadata_sigmoid=X_metadata_sigmoid,
clamp_min=clamp_min,
clamp_max=clamp_max,
callbacks=callbacks)
log.write('Finished training!')
if not use_hard_negative_mining:
# Not mining, so we can delete the training data early to save on resources
del train_dl, training_data
training_data = None
train_dl = None
# Checking how many epochs have actually passed. If more than 100, the fitted line will be printed!
epochs_passed = len(history)
include_line_fit = False
if epochs_passed > 100:
include_line_fit = True and include_line_fit_in_metrics_after_training
if writing_metrics_enabled:
log.write('Plotting and saving loss and acc plots...')
mil_metrics.write_history(history, history_keys, metrics_dir)
mil_metrics.plot_losses(history, metrics_dir, include_raw=True, include_tikz=True, clamp=2.0,
include_line_fit=include_line_fit)
mil_metrics.plot_accuracy_bags(history, metrics_dir, include_raw=True, include_tikz=True,
include_line_fit=include_line_fit)
mil_metrics.plot_accuracy_tiles(history, metrics_dir, include_raw=True, include_tikz=True,
include_line_fit=include_line_fit)
mil_metrics.plot_accuracies(history, metrics_dir, include_tikz=True, include_line_fit=include_line_fit)
mil_metrics.plot_dice_scores(history, metrics_dir, include_tikz=True, include_line_fit=include_line_fit)
mil_metrics.plot_sigmoid_scores(history, metrics_dir, include_tikz=True, include_line_fit=include_line_fit)
mil_metrics.plot_binary_roc_curves(history, metrics_dir, include_tikz=True)
if model.enable_attention:
mil_metrics.plot_attention_otsu_threshold(history, metrics_dir, label=1, include_tikz=True)
mil_metrics.plot_attention_entropy(history, metrics_dir, label=1, include_tikz=True)
mil_metrics.plot_attention_otsu_threshold(history, metrics_dir, label=0, include_tikz=True)
mil_metrics.plot_attention_entropy(history, metrics_dir, label=0, include_tikz=True)
del include_line_fit
##################
# TESTING START
##################
if testing_model_enabled:
log.write('Testing best model on validation and test data to determine performance')
test_dir = out_dir + 'metrics' + os.sep + 'performance-validation-data' + os.sep
test_model(model, model_save_path_best, model_optimizer, data_loader=validation_dl, out_dir=test_dir,
bag_names=bag_names, X_raw=X_raw, y_tiles=y_tiles)
if data_split_percentage_test > 0:
test_dir = out_dir + 'metrics' + os.sep + 'performance-test-data' + os.sep
test_model(model, model_save_path_best, model_optimizer, data_loader=test_dl, out_dir=test_dir, X_raw=X_raw,
bag_names=bag_names, y_tiles=y_tiles)
########################
# HARD NEGATIVE MINING
########################
if use_hard_negative_mining:
log.write('[0/4] Hard Negative Mining: Pre-Processing')
# Hard Negative Mining
# train_dl.transform_enabled = False
log.write('[1/4] Hard Negative Mining: Finding false positives')
false_positive_bags, attention_weights_list, false_positive_bags_raw = omnisphero_mining.get_false_positive_bags(
trained_model=model,
train_dl=train_dl,
X_raw=X_raw)
log.write('[2/4] Hard Negative Mining: Finding hard negatives')
hard_negative_instances, hard_negative_instances_raw = omnisphero_mining.determine_hard_negative_instances(
false_positive_bags=false_positive_bags, attention_weights=attention_weights_list,
false_positive_bags_raw=false_positive_bags_raw,
magnitude=hnm_magnitude)
if not len(hard_negative_instances):
log.write('[?/?] Hard Negative Mining: No hard negative instances found!')
return
log.write('[3/4] Hard Negative Mining: Creating new bags')
n_clusters = math.ceil(len(training_data) * hnm_new_bag_percentage + 1)
new_bags, new_bags_raw, new_bag_names = omnisphero_mining.new_bag_generation(hard_negative_instances,
training_data,
hard_negative_instances_raw=hard_negative_instances_raw,
n_clusters=n_clusters)
log.write('[4/4] Hard Negative Mining: Adding new bags to the dataset')
training_data, X_raw, bag_names = omnisphero_mining.add_back_to_dataset(training_ds=training_data,
new_bags=new_bags,
new_bag_names=new_bag_names,
X_raw=X_raw, new_bags_raw=new_bags_raw,
bag_names=bag_names)
# Fitting a new model with the mined bags
if augment_train:
train_dl = OmniSpheroAugmentedDataLoader(training_data, batch_size=1, shuffle=shuffle_data_loaders,
transform_enabled=augment_train, transform_data_saver=False,
**loader_kwargs)
else:
train_dl = DataLoader(training_data, batch_size=1, shuffle=shuffle_data_loaders, **loader_kwargs)
mined_out_dir = out_dir + os.sep + 'hnm' + os.sep
os.makedirs(mined_out_dir, exist_ok=True)
epochs = math.ceil(epochs * 1.5)
f = open(mined_out_dir + 'mining.txt', 'w')
f.write('Hard Negative Mining parameters:')
f.write('\nMining dir: ' + mined_out_dir)
f.write('\nEpochs: ' + str(epochs))
f.write('\nTraining data training bag mult: ' + str(hnm_new_bag_percentage))
f.write('\nTraining data new bag count: ' + str(n_clusters))
f.close()
del f
# Saving new bags to disk
sample_preview.save_hnm_bags(mined_out_dir + 'bags' + os.sep, new_bags, new_bags_raw, new_bag_names)
print('Fitting a new model using HNM bags!')
history, history_keys, model_save_path_best = models.fit(model=model,
optimizer=model_optimizer,
epochs=epochs,
training_data=train_dl,
validation_data=validation_dl,
out_dir_base=mined_out_dir,
data_loader_sigmoid=data_loader_sigmoid,
X_metadata_sigmoid=X_metadata_sigmoid,
sigmoid_evaluation_enabled=sigmoid_evaluation_enabled,
save_sigmoid_plot_interval=save_sigmoid_plot_interval,
checkpoint_interval=None,
sigmoid_video_render_enabled=sigmoid_video_render_enabled,
render_fps=render_fps,
clamp_min=clamp_min,
clamp_max=clamp_max,
bag_names=None,
# TODO add bag names
callbacks=hnm_callbacks)
# Plotting HNM metrics
log.write('Plotting HNM and saving loss and acc plots...')
metrics_dir = mined_out_dir + os.sep + 'metrics' + os.sep
os.makedirs(metrics_dir, exist_ok=True)
epochs_passed = len(history)
include_line_fit = False
if epochs_passed > 100:
include_line_fit = True
mil_metrics.write_history(history, history_keys, metrics_dir)
mil_metrics.plot_losses(history, metrics_dir, include_raw=True, include_tikz=True, clamp=2.0,
include_line_fit=include_line_fit)
mil_metrics.plot_accuracy_bags(history, metrics_dir, include_raw=True, include_tikz=True,
include_line_fit=include_line_fit)
mil_metrics.plot_accuracy_tiles(history, metrics_dir, include_raw=True, include_tikz=True,
include_line_fit=include_line_fit)
mil_metrics.plot_accuracies(history, metrics_dir, include_tikz=True, include_line_fit=include_line_fit)
mil_metrics.plot_dice_scores(history, metrics_dir, include_tikz=True, include_line_fit=include_line_fit)
mil_metrics.plot_sigmoid_scores(history, metrics_dir, include_tikz=True, include_line_fit=include_line_fit)
mil_metrics.plot_binary_roc_curves(history, metrics_dir, include_tikz=True)
if model.enable_attention:
mil_metrics.plot_attention_otsu_threshold(history, metrics_dir, label=1, include_tikz=True)
mil_metrics.plot_attention_entropy(history, metrics_dir, label=1, include_tikz=True)
mil_metrics.plot_attention_otsu_threshold(history, metrics_dir, label=0, include_tikz=True)
mil_metrics.plot_attention_entropy(history, metrics_dir, label=0, include_tikz=True)
del include_line_fit
# Testing HNM models on test data
log.write('Testing HNM best model on validation and test data to determine performance')
test_dir = mined_out_dir + 'metrics' + os.sep + 'performance-validation-data' + os.sep
test_model(model, model_save_path_best, model_optimizer, data_loader=validation_dl, out_dir=test_dir,
bag_names=bag_names, X_raw=X_raw, y_tiles=y_tiles)
if data_split_percentage_test > 0:
test_dir = mined_out_dir + 'metrics' + os.sep + 'performance-test-data' + os.sep
test_model(model, model_save_path_best, model_optimizer, data_loader=test_dl, out_dir=test_dir, X_raw=X_raw,
bag_names=bag_names, y_tiles=y_tiles)
# Finished HARD NEGATIVE MINING at this point
##################################
# FINISHED TRAINING - Cleaning up
##################################
log.write("Finished training and testing for this run. Job's done!")
del train_dl, training_data
del validation_dl, test_dl, X_raw, y_tiles
############################
# PREDICTING THE TEST DATA
############################
prediction_paths: [[str]] = []